Multilingual version

[ZodiacFRA]

Hi :) thanks for your repo it's quite useful !

How much work do you think it would take me to make this work with the multilingual version?

I haven't seen any .onnx weights for the multilingual version on their hf page. Do you think most of the work is converting the .safetensors to .onnx?

Have a great day :)

[DDATT]

Hi there,

Thanks for the kind words! I'm glad to hear you're finding the repo useful.

And yes, you are exactly right. Both the turbo and multilingual versions share the same pipeline and have nearly identical model architectures. Because of this, you should be able to reuse most of the current C++ codebase for the inference part.

But, since resemble-ai hasn't published their script for exporting the Hugging Face model to ONNX, you'll likely need to spend a significant amount of time figuring out the conversion yourself. It's very common that some PyTorch layers in the model aren't natively compatible with ONNX export, which means you might have to manually adjust or rewrite those layers to make the export work (I experienced this myself when converting StyleTTS2 from PyTorch to ONNX lol).

Have a great day too!